research · NYU WIRELESS
Spatial Intelligence
Runtime object association when the camera moves — and when the pose you trusted is wrong.
Problem framing
Keep object identity and association coherent as viewpoint changes. Clean pose is the easy case. Injected pose noise is the honest case. I treat them as separate evaluation stories so a single headline number cannot hide a brittle pipeline.
Results
Coverage and pseudo-label quality are reported separately. A system can look strong on matched cases while quietly skipping hard ones, or while training on labels that are not as clean as the paper title implies.
Methodology
- Visual, geometric, and relational cues all participate — none are treated as sufficient alone.
- Leakage controls and ablations are part of the research, not afterthoughts.
- Pose noise is a first-class evaluation condition, not a footnote.
The implementation repository is private. Model weights, dataset dumps, training configs, and the private repo are not published. This page describes the research question and the standard of evidence — enough to start a conversation.